← All stories
August 15, 2026

From Pulpit to Page

By Jonathan Pennington

How a sermon preached on Sunday becomes a published, described, topic-tagged, cross-linked web page, with almost nothing typed by hand.

The pipeline

1

The sermon is preached and uploaded to SermonAudio

The sermon is preached, streamed live, and uploaded to SermonAudio. SermonAudio then automatically generates a full text transcript of every word of the sermon available on its backend.

2

The website notices on its own

A background service on the church's server polls the SermonAudio API once an hour. When a new sermon appears, it pulls the title, preacher, date, audio, video, and transcript. If the transcript or video isn't ready yet, it rechecks until it is.

3

Claude reads the transcript and drafts the page

The full transcript goes to Claude over the API: Opus 5 at max effort. From the actual preaching transcript, it drafts three things: a description, a key Bible verse, and topic tags chosen from a fixed 26-term vocabulary (eternal security, Holy Spirit, prayer, the blood, and so on). Free-form tags are rejected; only the controlled list survives.

4

A human decides, modifies, or delegates

The import lands in a Pending Review queue on the admin dashboard, invisible to the public, where it can be edited and approved. Two switches control the machine: Auto-import pauses the whole watcher, and Auto-publish lets a fully successful draft go live with no waiting, while any import whose draft came back incomplete always stays held regardless.

5

The page publishes itself into the archive

The sermon gets a readable web address from its title, its metadata is injected for search engines, it joins the sitemap and podcast feed, and the page quotes its passage's actual KJV text. No webmaster required.

The guardrail that makes it trustworthy

The AI is never allowed to quote scripture. It may only name references, and a reference reaches a page only if it resolves against the church's own database of all 31,102 King James verses.

Verse text on the site is always rendered from that table, never from model output. Structurally, no AI-paraphrased or non-KJV Bible text can ever appear. The King James position is enforced by architecture, not by trust.

The resolver is forgiving about how a reference is written ("Psalm 103:1", "1st Samuel") but unforgiving about existence: if Claude names a passage that doesn't parse against the real Bible, the reference is dropped.

Healing the back catalog

Before this pipeline existed, months of sermons were typed into the dashboard by hand, many with pasted verse text standing in for a description. The watcher repairs those too: when a hand-entered sermon's SermonAudio counterpart shows up, it is matched rather than re-imported. Matching is deliberately conservative (same preach date, agreeing titles, no contradicting service time), and an ambiguous match is logged and skipped, never guessed. A matched sermon keeps its address and title but gains the transcript and a fresh Claude-drafted description, passage, key verse, and topics. Roughly 160 sermons in the archive have been tagged this way.

"Keep Listening": related sermons, ranked by substance

Every sermon page surfaces up to three related sermons. Not "recent posts," but preaching that genuinely overlaps, scored by what the sermons are actually about:

+4expounds the same book of the Bible
up to +3nearness of chapter within that book (adjacent chapters score highest)
+3 eachshared topic tag from the 26-term vocabulary
+1same preacher

A sermon on Romans 8 pulls other Romans preaching first, then anything tagged with its themes. Because the tags come from full transcripts rather than titles, the connections reflect what was actually preached. The links also weave the archive together for search engines: every page vouches for its neighbors.

And it costs about a nickel a sermon

5–7¢ measured per Opus 5 draftA few dollars tagged the entire back catalogOne switch pauses all spendUnder $100 / year in total hosting costs

The AI credits are pennies. In a side-by-side test on two real sermon transcripts, an Opus 5 draft covering the description, key verse, and topics together cost 4.5 cents on one sermon and 7.2 cents on the other. At a few sermons a week, the pipeline's entire AI bill is a few dimes a month. If it ever needed to stop, one admin dashboard switch pauses the machine. Combined with the $10 a year domain registration and a $6/month Hetzner VPS for hosting, the entire site costs less than $100 a year to run and fully maintain.

What a visitor actually gets

A sermon page with the preaching embedded, a real description of what was preached, the key passage quoted in the King James Bible from the church's own verse table, and three genuinely related sermons to continue with. Google gets the same page as structured data: validated video markup, clean addresses, and a feed that updates the moment a sermon publishes.

The preacher preaches, and everything else, the page, the description, the tags, the scripture quotation, the related preaching, happens on its own, with the King James text guarded at every step. We don't know of another church running anything like it.